Improve Eviction Policy

[Tookmund]

(Copied from http://pqxx.org/development/swapspace/ticket/11)

Work on the algorithm that chooses which swapfile to deallocate first. Current algorithm picks largest active swapfile that doesn't bring it below the target state.

Other options might be:

• Free file with least used pages to minimize "swapback" activity.
• Look for "best fit," i.e. plan a combination of swapfiles that can be freed to get as close as possible to the target state. Assuming nothing else changes in the meantime, the current scheme may actually be an efficient, greedy implementation of this.
• Try to balance swapback load and approximation of target state.

[wmertens]

I think minimal swapback makes most sense, especially if you always allocate in e.g. 500MB chunks.

I actually don't understand why the swapfiles are dynamically sized, is there a benefit to that?

[Tookmund]

Quoting from docs/comparison:

There's one twist to other swap managers: swapspace adapts the sizes of the swap files it creates. The Linux kernel may limit the total number of swap files, but as you need more, the daemon will create larger and larger ones. That way, creating your first swap file only takes a few seconds and won't slow down your system much. As you need more virtual memory, it will start to take longer but also give you more space--so it happens less often.

[rubo77]

The tickets are in the archive: https://web.archive.org/web/20160915022004/http://pqxx.org/development/swapspace/report/3